Written by Julien Ricciarelli-Bonnal
9 October 2026
The Essentials
OpenAI is reportedly running at roughly $50 billion in annualised revenue, even though a figure closer to $70 billion had circulated only weeks earlier. The company has not suddenly lost $20 billion in sales. Much of the difference comes from the way certain transactions through cloud partners are accounted for and from an adjusted figure previously used to make OpenAI easier to compare with Anthropic. The episode reveals something broader about the AI economy: once revenues, investments and valuations reach extraordinary levels, understanding exactly what a number measures becomes almost as important as the number itself.
Has OpenAI somehow lost $20 billion in revenue within a matter of weeks? Framed that way, the story would suggest a dramatic deterioration in one of the fastest-growing technology businesses in the world. In September, a figure approaching $70 billion in annualised revenue was circulating around the company. In early October, Axios reported that OpenAI had instead told investors that its annualised revenue was running closer to $50 billion.
The instinctive response is to subtract one number from the other and look for an explanation for the missing billions. That would be misleading. The two figures were not constructed in exactly the same way, even though they have sometimes circulated as if they were directly comparable.
According to Axios, the higher figure was an adjusted estimate designed in part to make OpenAI easier to compare with Anthropic. A significant part of the difference comes from how revenue generated through cloud partners is recorded. OpenAI has not therefore gone from $70 billion to $50 billion because customers suddenly disappeared. What changed is largely the accounting lens through which part of the business was being presented.
That may sound like a technical detail. When the adjustment itself is worth roughly $20 billion, it stops feeling like one.
Even $50 billion does not mean OpenAI collected $50 billion over the past twelve months
Before looking at the difference between $50 billion and $70 billion, another distinction matters: these are annualised revenues.
Annualisation takes a recent level of activity and projects what it would represent over a full year if the same pace continued. If a business generated $4 billion in one month, for example, that pace could be described as roughly $48 billion on an annualised basis. It does not mean the company actually collected $48 billion over the previous twelve months.
This type of measure is particularly useful for companies growing very quickly because historical revenue can become outdated almost as soon as it is reported. Looking only at the previous twelve months may substantially understate the scale a business has already reached by the end of the period.
Annualised revenue is therefore not inherently misleading. It simply answers a different question. Problems begin when “$50 billion in annualised revenue” gradually becomes “OpenAI makes $50 billion a year” in headlines, conversations and market comparisons, particularly when the number is then placed beside another figure calculated under different assumptions.
This matters especially in artificial intelligence, where growth can change dramatically over a few quarters and where the size of the numbers has itself become part of the story. The sector has already reached a stage in which acquisitions, infrastructure programmes and commercial agreements routinely involve sums that would once have defined entire technology industries. Stripe’s reported $8 billion acquisition of OpenRouter was already a reminder that enormous amounts of value are now being assigned not only to AI models, but to the infrastructure surrounding them.
The larger the numbers become, the less room there is for treating their definitions as a footnote.
The gap between $50 billion and $70 billion is partly about a very old accounting question: who records the sale?
The mechanism described by Axios is relatively straightforward in principle. When an artificial intelligence service is sold through a cloud partner, several companies may participate in the same transaction. The accounting question becomes how much of the customer’s spending should appear as revenue for the model provider itself.
Anthropic reportedly recognises some sales made through cloud partners on a gross basis. If a customer spends $100 accessing an AI service through a cloud platform, the full $100 may appear as revenue before the partner’s share is recorded elsewhere as a cost.
OpenAI uses a different presentation for some of these transactions and records only the portion that ultimately belongs to it. The customer may have spent exactly the same amount and the underlying economic activity may be comparable, yet the number appearing at the top of the income statement can look very different.
Axios reports that the figure approaching $70 billion associated with OpenAI in September had been adjusted to show what its business might look like under a treatment more comparable with Anthropic’s. The approximately $50 billion figure is closer to the way OpenAI normally measures its own annualised revenue.
This does not automatically mean that one company is counting honestly while another is inflating its performance. Different accounting treatments can be legitimate depending on the contractual structure, who controls the service, who owns the customer relationship and who carries responsibility for delivering the product.
The more interesting problem begins when an adjusted comparison figure escapes from the context in which it was created and starts circulating as if it were simply another version of reported revenue.
In the AI economy, the most spectacular number travels fastest
Artificial intelligence has entered an unusual financial environment. Infrastructure investments are measured in tens or hundreds of billions of dollars, company valuations can change dramatically between funding rounds, data centre announcements resemble industrial policy and every leading company is under pressure to demonstrate that extraordinary spending is being matched by equally extraordinary growth.
Under those conditions, numbers are not merely accounting information. They become part of the competitive narrative.
There is an obvious psychological difference between saying that a company is approaching $50 billion in annualised revenue and saying that it is approaching $70 billion, even if the underlying commercial activity has barely changed. Both figures may be defensible when their methodology is explained. Without that explanation, however, they create very different perceptions of momentum, scale and competitive position.
The issue extends far beyond OpenAI. Private technology companies do not publish financial statements with the same regularity and level of detail as listed corporations. Information often reaches the market through investor documents, fundraising discussions, internal presentations, partners or selected disclosures to journalists. Meanwhile, expressions such as annualised revenue, bookings, contract value, gross merchandise volume and adjusted revenue can all describe different aspects of the same business.
None of these measures is necessarily illegitimate. But once they begin appearing beside one another in headlines, the distinction between measuring performance and narrating performance can become surprisingly thin.
The AI sector makes that tension particularly visible because the underlying numbers are now so large. A methodological adjustment that would once have represented a few million dollars can now create an apparent gap equivalent to the annual turnover of a major multinational company.
Comparing OpenAI with Anthropic now requires comparing their definitions as well
The irony is that the adjustment was reportedly intended to improve comparison. Investors naturally want to understand how OpenAI and Anthropic perform relative to one another, and that becomes difficult if both companies treat some commercial flows differently.
Normalising the numbers can therefore be perfectly reasonable. Financial analysts routinely adjust reported figures in order to compare businesses on a more consistent basis.
The paradox is that every adjustment introduces another layer between the raw economic activity and the number ultimately quoted. A figure created to make two companies easier to compare can later be mistaken for the company’s own reported revenue, particularly once it begins travelling without the methodological explanation attached to it.
A serious comparison therefore requires much less glamorous questions than “Who has more revenue?” Are we looking at revenue already recognised or at an annualised run rate? Are partner sales recorded gross or net? Are future contracted revenues being included? Do the figures cover the same period? Are costs associated with those revenues treated consistently?
The same discipline is needed when discussing profitability. A company can describe itself as profitable according to one operational measure while excluding significant categories of expense from that calculation. That does not automatically make the measure useless. It simply means that terms such as revenue, profit and margin become increasingly unhelpful when detached from the definitions used to construct them.
Complexity should not lead us to distrust every financial number. It should lead us to read them more carefully.
The real question may not be whether OpenAI is at $50 billion or $70 billion
The gap makes for an excellent headline because it appears that $20 billion has vanished. The more interesting conclusion is almost the opposite: nothing needs to vanish for an enormous numerical difference to appear.
Artificial intelligence has reached a scale where financial conventions can materially alter the public perception of an industry. For years, the leading companies were mostly compared through model performance, user growth, benchmark scores and technological capabilities. They are now becoming large enough that revenue recognition, infrastructure costs, margins and the quality of those revenues will matter just as much.
That transition also applies to the companies adopting the technology. AI was initially evaluated through impressive demonstrations: how quickly a model could generate content, how many tasks could be automated or how sophisticated an assistant appeared. As the technology becomes embedded in actual organisations, the questions become much less theatrical. What value is being created? How is it being measured? What does it cost to produce that value, and what would have happened without the technology?
That is precisely why AI Business Governance increasingly requires companies to separate attractive metrics from the uses that genuinely produce measurable business value. A number can be accurate and still be unhelpful if nobody has defined what it is supposed to represent.
OpenAI has therefore probably not “lost” $20 billion in revenue. What disappeared along the way was part of the definition attached to the number as it moved from an investor comparison into the wider conversation.
That may be the more useful lesson. When the AI economy was measured in millions, methodological differences could look like technical details. When it is measured in tens of billions, the footnote can itself be worth the size of a multinational company.
If your company is trying to separate AI promises from value that can actually be measured, we can help you structure the use cases, priorities and indicators that genuinely matter.
Written by Julien Ricciarelli-Bonnal
9 October 2026
